Employing deep learning framework for improving solar panel defects using drone imagery

Abstract

This research describes a unique method for identifying and categorizing solar panel problems using RGB and thermal pictures captured by drones. The first step of the suggested technique is to identify solar panels in the photographs by using a CNN based on YOLOv5 architecture model that was trained and tested on an annotated dataset of solar panels. A number of computer vision techniques were used to separate the panels from their backgrounds in order to get around the accuracy issues with the detector. The panels were then classified as normal or anomalous using a state-of-the-art EfficientNet classifier, which was trained on a synthetic dataset. The anomalies were then divided into four categories: cell, multi-cell, diode, and multi-diode. The results obtained from this research demonstrate the viability and potential of employing drones to identify and categorize solar panel problems and emphasize the significance of creating precise models to enhance solar park management.